This repository provides a comprehensive course on Occupancy Networks for autonomous driving perception, targeting researchers and engineers. It aims to demystify the complex field of 3D semantic occupancy perception, offering a structured learning path from foundational concepts to practical deployment, with the benefit of enabling more robust and accurate environmental understanding for autonomous vehicles.

How It Works

The course covers a wide spectrum of Occupancy Network approaches, including pure vision-based methods (e.g., TPVFormer, OccFormer, OccNeRF) and multi-sensor fusion techniques leveraging point clouds. It delves into the underlying principles, architectural designs, loss functions, and experimental methodologies of various state-of-the-art algorithms, facilitating a deep understanding of their strengths and weaknesses.

Quick Start & Requirements

• Install/Run: Clone the repository and use the provided Docker script:

  git clone https://github.com/Charmve/OccNet-Course --recursive
  cd OccNet-Course
  ./scripts/start_dev_docker.sh
  ./scripts/goto_dev_docker.sh bash docker/run_after_start_docker.sh
  

• Prerequisites: Docker, NVIDIA GPU with CUDA support.
• Resources: Docker image setup and execution.
• Links: Online Course Website

Highlighted Details

• Comprehensive coverage of Occupancy Network algorithms, from BEV perception to 3D occupancy.
• Detailed explanations of both pure vision and multi-modal (point cloud) approaches.
• Practical code implementation and deployment guidance, including model quantization for edge devices (NVIDIA, Horizon J5).
• Covers key datasets (nuScenes, SemanticKITTI) and benchmarks in the field.

Maintenance & Community

The project is actively updated, with a changelog available. A WeChat group is available for Q&A (contact: Yida_Zhang2). The author is an experienced autonomous driving engineer.

Licensing & Compatibility

The repository's license is not explicitly stated in the README. Code examples and benchmarks may be subject to their original licenses. Compatibility for commercial use or closed-source linking would require further investigation into the specific licenses of included algorithms and datasets.

Limitations & Caveats

The README advises against forking due to ongoing updates. While the course aims for comprehensiveness, specific details on the exact state of the "standard version" code release (expected April 2024) are not fully detailed. The primary online course website is independently hosted and may have availability limitations.